If you control the machines, then it's actually pretty easy - you simply alter the votes of people who actually voted. There's no paper receipt or general mechanism for voters to verify they voted how they think they did. And it's absent for a good reason, which is that a nonrepudiatable vote receipt opens elections to coercion. A malicious machine could use a registration database (available to campaigns) to target elderly or otherwise likely tech unsavvy voters - people, who even if they found out their vote was altered, would probably have their complaints dismissed as user error like butterfly ballots in 2000.

If your manipulation is spread across counties, and flips what would have already been a close election, it would almost certainly be undetected. Even if it was detected, the winning party (who doesn't even have to be in on the con) will defend the result. And without ironclad evidence of malfeasance, courts will be hesitant to give the appearance of overturning an election result.

It's critical that we have real security in our elections.

  a priori knowledge of who is registered but will not vote

You can indentify large numbers of such candidates with high accuracy even without complex programming, let alone anything approaching AI, and the number of prospects can be huge in registrars which don't purge dead/moved voters. There is often an order of magnitude more unvoted ballots than the vote margin for most offices and measures.

And generally, the party central committees get both registration data and turnout data for the entire county for free.

Like the scene from gangs of new York? Where they vote, shave, vote again...